Programmable photonic neural networks combining WDM with coherent linear optics

Abstract

Neuromorphic photonics has relied so far either solely on coherent or Wavelength-Division-Multiplexing (WDM) designs for enabling dot-product or vector-by-matrix multiplication, which has led to an impressive variety of architectures. Here, we go a step further and employ WDM for enriching the layout with parallelization capabilities across fan-in and/or weighting stages instead of serving the computational purpose and present, for the first time, a neuron architecture that combines coherent optics with WDM towards a multifunctional programmable neural network platform. Our reconfigurable platform accommodates four different operational modes over the same photonic hardware, supporting multi-layer, convolutional, fully-connected and power-saving layers. We validate mathematically the successful performance along all four operational modes, taking into account crosstalk, channel spacing and spectral dependence of the critical optical elements, concluding to a reliable operation with MAC relative error $$< 2\%$$ < 2 % .

WDM equipped universal linear optics for programmable neuromorphic photonic processors

“…We assume all modulators are realized in Si technology and exhibit wavelength dependent behavior when shared among multiple channels. This effect is compensated by PSs in the bias branch, as detailed in [38]. Moreover, we introduce an additional variable optical attenuator in the bias branch which matches the excess attenuation seen by the signal in the OLAU and guarantees the appropriate bias signal level for conversion of the sign from the phase to the field magnitude.…”

“…Expanding the single-neuron single-wavelength OLAU toward multi-neuron layouts that can also support WDM operation can open completely new parallelization perspectives and pave the road toward TMAC/sec/axon compute rates, indicated by the green bullet in figure 1. The WDM-enabling credentials of the OLAU have been already highlighted in our recent work on multichannel coherent photonic neurons [38], where the combined use of WDM with an OLAU and simple optical switching elements has been shown to create a unique programmable photonic neural network (PNN) setup, where the same hardware can be used for fully-connected, convolutional or multi-neuron NN layers. This work has demonstrated that wavelength dependence of the underlying photonic components plays marginal role as long as amplitude and phase matching is done in the bias branch, theoretically limiting the number of employed optical channels to the ratio of the bandwidth of the de/multiplexers (DE/MUXes) and channel spacing.…”

Massively Scalable Wavelength Diverse Integrated Photonic Linear Neuron

“…For sufficiently small radii MRD (or, conversely, sufficiently large freespectral-range (FSR)), we can place additional MRDs at new resonant wavelengths (or along specific channels, similar to wavelength-division-multiplexing (WDM) in optical communications). Recently, a similar WDM scheme was proposed using MZMs, but the WDIPLN implementation with ring resonators have the advantage that the wavelength multiplexing/de-multiplexing is done intrinsically using the single device in a compact footprint [18]. We show a schematic for the naïve WDIPLN in Figure 2(a).…”

“…Carefully designed MZI mesh circuits can implement any real-valued matrix on-chip by reconfiguring phase shifters according to singular value decomposition [1,16]. The coherent optical linear neuron (COLN) architecture implements vector-vector multiplication and summation by leveraging MZIs as inputs and weights for multiplication and coherent interference as the summation operator [17,18]. Wavelength incoherent designs rely on multi-wavelength operation, often employing a broadband photodetector to perform destructive summation of the various signals.…”

Massively scalable wavelength diverse integrated photonic linear neuron

“…Carefully designed MZI mesh circuits can implement any real-valued matrix on-chip by reconfiguring phase shifters according to singular value decomposition [1,16]. The coherent optical linear neuron (COLN) architecture implements vector-vector multiplication and summation by leveraging MZIs as inputs and weights for multiplication and coherent interference as the summation operator [17,18]. Wavelength incoherent designs rely on multi-wavelength operation, often employing a broadband photodetector to perform summation, but through this method the phase of the signal is lost to optical-electrical (OE) conversion.…”